77

APPLICATION OF COLLECTIVE RISK THEORY TO ESTIMATE

VARIABILITY IN LOSS RESERVES

ROGER M. HAYNE

Abstract

The intent of this paper is to present un introduction to

Collective Risk Theory for the first time reader and consid-
erations in applying that theory to estimate variability in loss
reserves. It begins with u brief introduction to the basic con-
cepts of Collective Risk Theory along with a survey of some
of the techniques developed to date to estimate the aggregate
distribution of losses. With this framework, descriptions of
some applications to loss reserves are discussed, with atten-
tion paid to the assumptions inherent in those methods and
some problems that arise in applying this theory to reserves.
Of note are questions that are not directly addressed b>l this
model; in purticular, parumeter uncertainty. Included are ref-
erences which, it is hoped, will lead the interested reader
.further into the applications to date.

1. INTRODUCTION

The question of the amount of variability inherent in loss reserve

estimates has gained more notice in recent years. In fact, Principles 3
and 4 of the Statement of Principles Regarding Property and Casualty
Loss and Loss Adjustment Expense Reserves [I] state,
“3. The uncertainty inherent in the estimation of required provisions for unpaid
losses or loss adjustment expenses implies that a range of reserves can be
actuarially sound. The true value of the liability for losses or loss adjustment
expenses at any accounting date can be known only when all attendant
claims have been settled.
4. The most appropriate reserve within a range of actuarially sound estimates
depends on both the relative likelihood of estimates within the range and
the financial reporting context in which the reserve will be presented.”

Quantification of the variability in reserve estimates will thus be

useful in the determination mentioned in Principle 4. In addition, knowl-
78 VARlABlLITY IN 1,OSS RESERVES

edge of the statistical distribution of reserves is also useful in discussing

the impact of reserve discounting on insurer capacity and solidity. One
author has already cited this as a favorable result of discounting in that
discounting of reserves would “increase the statutory capacity of the
insurance industry. Statutory surplus would increase as loss reserve lia-
bilities were reduced [2].” However, simply discounting reserves will
not necessarily increase financial strength or capacity. Rather, a better
measure of that capacity is probably the ability of surplus to protect
solvency. Without knowledge of the variability of the reserve estimates,
the assessment of the strength of a company at a given level of surplus,
and hence capacity, probably cannot be made accurately.
There are several techniques which are available to assessthe finan-
cial solidity of a given amount of surplus. Methods that have been
advanced for this purpose include “confidence limit” approaches, Ruin
Theory, and Utility Theory, along with a rather comprehensive model of
the operations of an insurer (see [3] and [4] for this latter application).
In each case, however, their application requires an estimate of the
statistical distribution of the reserves.
The intent of this paper is to discuss the framework of Collective
Risk Theory as one approach that can be used to estimate the statistical
distribution of reserves. No prior exposure to Collective Risk Theory is
assumed; however, it is hoped that the references will provide a good
starting place for the reader who wants to pursue this subject further.

2. THE COLLECTIVE RISK MODEL

The basic collective risk model approaches the question of the dis-
tribution of total reserves by modeling the claim process faced by an
insurer. It considers the interaction between the distribution of the number
of claims and the distribution(s) of the individual claims by calculating
loss (or reserve) T as the sum
T = X, + Xz + . . . + X,v, (2.1)
where the number of claims N is randomly selected. and each of the
claims XI, XZ, . . , X,%,is randomly selected from claim size distribu-
tion(s).
There is a significant amount of literature which addresses this model
and its applications to casualty insurance. The primary source is probably
the text by Beard, Pentikainen and Pesonen 151. Other complete texts
 VARIABILITY IN LOSS RESERVES 79

dealing with Collective Risk Theory and its applications are those by
Borch [6], Biihlmann [7] and Seal [8]. The papers by Borch [9] and
Pentikainen [lo] also consider this model from a fairly broad point of
view.
There are some useful properties of the distribution T under rather
broad assumptions. In particular, if
1. The number of claims N has moments
v =E(N)
vi = E[(N - v)‘] for i = 2, 3, and 4;
2. All claims are drawn from the same population with moments
x = E(X)
xi = E[(X - x)~] for i = 2, 3, and 4; and
3. All claims X and the number of claims N are all independent, then
the first four moments of the random variable T exist and are
given by
E(T) = vx (2.2)
E({T - E(r)}*] = x2v + x*v2 (2.3)
E[{T - J3T)131= x3v + 3X2XV2+ x3v,3 (2.4)
E[{T - E(~3)41= X4V + 3X2*(V2 - V + V*) + ‘tXX3vz +

~X*XZ(V~ + VV~) + x4v4. (2.5)

Comparable formulae for higher moments can also be derived if the
corresponding moments of the claim count and size distributions exist.
The paper by Mayerson, Jones and Bowers [ 1 l] gives a derivation of
these formulae.
These facts can and should be used to test the reasonableness of any
approximation to the distribution of T. In fact, one of the methods used
to approximate that distribution relies on these relationships.

3. APPROXIMATIONS OF THE DISTRIBUTION OF T

There have been many approaches used in estimating the distribution

of T, given distributions for the number of claims N and the size of those
claims. These methods can be broadly grouped into 3 classes:
80 VARIABILITY IN LOSS KbS1-RVkS

1. Monte Carlo Simulation.

2. Approximate Distributions, and
3. Analytic Approximation.

Monte Curio Simulation

Probably the most flexible of these approaches is that of Monte Carlo

Simulation. The idea is simple and directly follows the basic Collective
Risk Model above. Simply stated. the Monte Carlo Simulation algorithm
is composed of five steps:
1. Randomly select the number of claims N from the claim count
distribution.
2. Randomly select N claims, XI, X2. . X,V, from the claim size
distribution.
3. Calculate one observation from the distribution of T by the sum
x, + x* + ... + Xh..
4. Repeat steps 1 through 3 “several” times.
5. Estimate the distribution of T using the points generated in this
manner.
Conceptually, there is no limit on the form of the claim count or size
distributions used in Monte Carlo Simulation. They can both be discrete
or the claim size distribution can be continuous. Simulation with de-
ductibles and/or per claim loss limitations can also be easily handled in
this framework. In addition, the combination of several lines of insurance
or accident years can also be accommodated without much difficulty.
There are, however, prices to pay. First, the answer to how many is
“several” in step 4 is not clear. Often a signifcant number of simulations
must be run to obtain a clear enough picture of the distribution of T to
be useful in applications. One technique. though admittedly “brute
force ,” is to compare the results of two sets of simulations, say each of
1,000 trials. If the resulting distributions are “close enough” for the task
at hand, the combined distribution could be used as an approximation.
If, however, they differ significantly, more trials may be indicated. The
moments of the simulated distributions should bc compared to the the-
oretically expected moments in (2.2) through (2.5) to see if the simulation
is sufficiently close.
Another practical consideration is how to simulate the random selec-
tions from the claim count and claim size distributions. Care should be
 VARIABILITY IN LOSS RESERVES 81

taken as to the representations of the distributions. Finite distributions,

such as those based solely on empirical data, implicitly have upper
bounds. Thus, unless those upper bounds are to be explicitly considered
in the model, some of the variability inherent in the underlying distri-
bution may be lost.
One solution to this difficulty could be to use analytic distributions,
such as the lognormal or Pareto, to estimate the distributions in the
“tail.” In this way the empirical data could be used, and yet some of the
potentially unlimited nature of some risks can be captured.
The process used to make the random selections from the claim size
and count distributions may not be obvious. Most computer software
packages do provide “random” number generators which correspond to
a uniform distribution. In addition, there are algorithms which allow for
selections from other distributions, either directly or from selections from
the uniform distribution. The very useful text A Guide fo Simulation by
Bratley, Fox and Schrage [ 121 includes some of these algorithms for a
number of statistical distributions. That text also includes listings of
computer programs to perform those calculations.
One final consideration regarding Monte Carlo Simulation is the cost
in computer time. Factors influencing this time include the complexity
of the model used, the expected number of claims E(N), the degree of
accuracy required, and the amount of dispersion in the claim size distri-
bution. Simulations involving a great number of expected claims will of
course take longer to run. Not immediately obvious, though, is the fact
that if the claim size distribution is dispersed (a large standard deviation
as compared to the mean), there will generally be a greater number of
simulations necessary to achieve a desired level of accuracy than if the
claim size distribution is less dispersed.

Appro,uimate Distributions

Another method used to estimate the distribution of T involves as-

suming a statistical distribution and then using the “known” moments of
T to select the parameters of that distribution. Probably falling into this
category is the Normal-Power, or NP Approximation. This approach is
described in Beard, Pentikainen and Pesonen [5] and used by Mayerson,
Jones and Bowers [ 1I] and by Patrik and John [ 131. Although relatively
easy to apply, it does not seem to be sufficiently skewed for many
casualty applications. However, caution should be taken in applying this
82 VARIABILI I Y IN LOSS HtSf.KVI,S

approach. It can easily yield misleading results, or even nonsense, if

misapplied, especially if the variable to be approximated differs markedly
from the normal distribution,
This approach considers a transformation of the variable T which is
hoped to be approximately normal. Although the transformation can be
carried out to include several moments of the distribution of T, the
application in [ I l] stops at the third moment with the formula:
ro = ml + PI-:() + n& - I )ih + nrj(z;: - 3=,,)/‘24 -
m:(2+5zo)/36, (3.1)
where z() represents the 1OOepercentile of a standard normal distribution
and f. represents the approximate 100~ percentile of the distribution of
T. Here
111
, = W-I
m; = E[{T - E(7-)}‘]
n.3 = E[{T - E(7-)}3]/m;
UL, = E[{T - E(7-)}4]/rn: - 3
Using formulae (2.2) through (2.5). the various moments of T can
be found from those of the claim count and size distributions. The various
percentiles of the aggregate distribution can then be approximated.
A similar approach is followed by Venter in [ 141. In that paper,
transformations of the Beta and Gamma distributions are suggested as
forms for the distribution of aggregate losses. The Gamma distribution
is also suggested by Beard, Pentikainen and Pesonen, [ 5, page 1211.
Again, matching of moments is used to estimate the parameters of the
distribution. Pentikainen [ 151, Lau [ 161 and Philbrick [ 171 also present
approaches based on distribution fitting.
The be&it of this approach is its relative simplicity and, once the
moments are calculated, the ease with which the percentiles of the
aggregate distribution can be approximated. It does require, however,
that the form of the distribution be assumed and there are no readily
available tests of how well the distribution used fits the actual distribution
of T.
 VARIABILITY IN LOSS RESERVES 83

Analytic Appro.uimation

A third category of approximations of the distribution of T attempts

to analytically calculate that distribution. This approach generally looks
at the distribution of T as the sum

F(t) = 2 P(N = n) F,,(t), (3.2)

n=o
where P(N = n) is the probability of n claims and F,,(r) is the probability
that the sum of n claims will be less than t. The functions F,,(r) can then
be calculated in terms of the probability density function of the individual
claim size distribution. In the discrete case, for example, if F(K) is given
by
F(100) = 0.60
F(300) = 1.00,
then F?(x) will be given by
F2(200) = 0.36
F2(400) = 0.84
F2(6OO) = 1.00.
Since there are only two outcomes of the original distribution, a loss
of 100 with probability .6 and a loss of 300 with probability .4, the only
possible outcomes for the sum are 200 (two losses at 100 each), 400
(one loss at 100 and one at 300), and 600 (two losses at 300 each). The
resulting distribution is called the convolution of the probability density
function (p.d.f.) underlying F with itself. More generally, in the contin-
uous case, if&) and R(J)) are p.d.f.‘s for independent random variables
X and Y, then the sum Z = X + Y has the p.d.f. given by

(f*g)(z) = lx f(x)&-ax, (3.3)

-r

which is called the convolution off and g. Similar to multiplication

define f*” iteratively by

f*n =f*f*(“-‘) for n = 1, 2, . .

84 VARIAB1I.I I’Y IN LOSS RF.SEKVES

Then F,,(s) can be written in terms off*‘* as

I
F,,(x) = ox f‘*” (z)dz. (3.4)
I
If now the p.d.f. of the claim size distribution isf(x), then, combining
(3.2) and (3.4), the p.d.f. underlying the distribution of T can be written
as

h(r) = c P(N = n)f*” (1). (3.5)

,I=o
These formulae hold under rather broad conditions which guarantee
that the sum converges and the various f*” (.w)exist. If one is willing to
place some restrictions on the distribution of claim counts N, then (3.5)
can be further simplified.
A common approach is to considcr the characteristic function (or
Fourier transform) of the probability density function of the claim size
distribution
C[.f‘](r) = E[exp(irX)j. (3.6)
where i is the imaginary unit. Under rather broad regularity and integra-
bility conditions onJ: this function exists and is “unique.” Thus, if the
characteristic function is known then, theoretically at least, the under-
lying distribution function can hc found. A useful property of the char-
acteristic function is that
clJ‘*glcf, = Cl,mKIgl(t) (3.7)
iff‘and s arc independent p.d.f.‘s. Thus, under conditions sufficient for
the sums to exist.

Clhl(t) = c P(N = 11)C’[,f‘]($. (3.8)

,1--o

If N is further assumed to have a Poisson distribution with mean \‘, i.e.,

P(N = II) = e “r-“/n! ,

then C(h](r) can be written as

C[h](r) = c TV“v”C[.f’](t)“/r~!
,, -0
 VARIABILITY IN LOSS RESERVES 85

which reduces to

C[hl(Q = exp~~~(WlO)- 1)). (3.9)

Note that C[h](r) is the moment-generating function of the Poisson

distribution evaluated at the natural logarithm of the characteristic func-
tion of the claim size distribution. Under suitable regularity conditions,
this result generalizes to other claim count distributions. That is, the
characteristic function of the aggregate distribution is the moment-gen-
erating function of the claim count distribution evaluated at the natural
logarithm of the characteristic function of the claim size distribution.

Heckman and Meyers [ 181 present an algorithm which “inverts” this

characteristic function. They only require that the probability density
function for the distribution of claims by size be a finite step function.
Since any (reasonable) probability density function can be approximated
as closely as desired by such step functions, conceptually the algorithm
they developed should be applicable in any situation.

In addition, they relax the above condition that the claim count
distribution be Poisson, with variance and mean equal. Their algorithm
also applies to the cases when that distribution is binomial (with variance
less than the mean) and negative binomial (with variance greater than
the mean). They include a provision for the uncertainty in parameter
estimates in the choices of the distributions.

Finally, computer code is provided for the algorithm. The algorithm

is computationally rather efficient and can easily be run on a microcom-
puter with a mathematical co-processor in a reasonable amount of time.
In short, Heckman and Meyers provide a very valuable tool to estimate
the distribution of T and, for a very wide range of cases, effectively
solve that problem.

Another approach to this problem was taken by Panjer [ 191 and by

Sundt and Jewel1 [20]. In the simplest case, assuming the claim count
distribution is Poisson and the p.d.f. of the claim size distribution is
discrete and evaluated at equally spaced points, there is a recursive
formula which leads to a direct calculation of the distribution of T. Work
continues in this area (for example, Willmot [21]).
86 VARIAHII.If’Y IN I OSS RtStRVtS

4. APPLICATIONS IN LOSS RESERVES

It is interesting to note that the majority of the references listed so

far either deal with Risk Theory on its own or in relationship to various
aspects of ratemaking. There have been some recent papers dealing with
risks and uncertainty in loss reserve estimates (see [3], 141, [22], [23],
and [24]), but we have been unable to find any which deal directly with
considerations which enter with the application of this model to the
estimation of variability in loss reserves.
The model of the insurance process provided by Collective Risk
Theory seems a natural tool to apply in evaluating the degree of uncer-
tainty in loss reserve estimates. If, for example, under the independence
hypotheses listed in Section 2, the distribution of open and IBNR claims
(N) is known and the distribution of the size of those claims (X) is also
known, the methods outlined in Section 3 all provide ways to estimate
the distribution of total reserves (T).
One approach used at this point takes the actuary’s best estimate of
ultimate claim counts and losses as an estimate of the expected number
of claims E(N) and average claim size E(X). Statistical distributions are
then selected for each of these quantities.
If the Poisson is chosen as a model of the claim count distribution,
then the only parameter to estimate is its mean. Other distributions, such
as the binomial and negative binomial, allow for the variance of N to
differ from its mean. These are “well behaved” and can be easily accom-
modated in the algorithm described in [ 181.
The claim size distribution is usually assumed to be more complex.
Common choices include the lognormal, Pareto, and a transformed
Gamma, among others. An ad iwc approach is to select the distribution
to be used, assume that its mean corresponds with the average claim
size derived by the actuary’s best estimate, and then select the other
parameter(s) either judgmentally or based on characteristics of the line
under evaluation. This may be all that can be done in situations where
data for further analysis is lacking. If sufficient data is available, how-
ever, the techniques described by Hogg and Klugman [25] provide pow-
erful tools to select the “proper” distributions.
To better model the distribution of reserves for an insurer or self-
insured, accident (report, or policy) years are often considered separately,
 VARIABILITY IN LOSS RESERVES 87

with separate distributions of claim counts and claim sizes for each year.
This has the benefit of preserving differences in relative maturity and
maintaining greater homogeneity of claims within each year. The distri-
bution of total reserves can be calculated using convolutions of the
distributions for individual years if the various years are assumed to be
stochastically independent. The algorithm in [ 181 allows for such con-
volution. One “short-cut” sometimes taken is to approximate the 9.51h
percentile, for example, of the distribution of total reserves by the sum
of the 95’h percentiles of the distributions of reserves for various accident
years. A bit of reflection leads to the conclusion that this assumes that
the various distributions are perfectly correlated with each other.
There are many possible approaches that can be used to estimate the
distributions and resulting reserve variability estimates. What follows
here is a discussion of only one possible approach.
This refinement considers the distribution of reserves for an accident
year as the combination of the distributions of reserves in three cate-
gories: case reserves, development reserves, and IBNR reserves. In this
discussion, we consider reserves for reopened claims in the IBNR cate-
gory. This approach allows closer modeling of the various components
of the reserves. These three components also have respectively increasing
uncertainty, summarized in the following table:

Counts Amounts

Case Reserves Certain Certain

Development Reserves Certain Uncertain
IBNR Reserves Uncertain Uncertain

Distributions for Reported Claim Sizes-One Approach

If we group the first two categories, the case and development re-
serves, then the statistical uncertainty lies only in the variation of claim
sizes, since the number of the claims is known. Given an estimate of
the claim size distribution, methods presented in Section 2 could be
applied to estimate the distribution of these reserves.
The current distribution of open and reported claims may provide
some knowledge of this distribution. For more mature years, one could
consider the relationship between the distribution of claims at this stage
88 VARIAHII IIY IN LOSS KI:SI:R\ lb5

of development with the “ultimate” distribution of those same claims

and incorporate it. with the current distribution. to estimate the ultimate
distribution of claims.
As an example of one possible approach, Ict us assume that the
lognormal is an appropriate model for the distribution of X, the claim
size random variable. Then Z = In(X) has a normal distribution, and the
lognormal can be completely parameterized by the mean m and variance
.Y’ of Z. We select this parameterization for the distribution of X.
It then follows (see, for example, p. 38 of [26J) that maximum
likelihood estimators for m and ,s2arc obtained from the sample mean
and variance of the values ln(X,) where X, arc observed claims. As in
the normal case, the sample variance. using the number of sample points
as the denominator, is a biased estimate for .v2:therefore, a denominator
of jr - I is used to estimate s’.
Suppose, for example. that we arc trying to estimate the claim size
distribution for open and reported claims for accident year 1981 as of
December 3 1, 1988. That accident year is currently 84 months from the
beginning of I98 I.
We can calculate the estimators /?rxJand sft of the VI and s2 parameters
for reported claims for “mature” accident years at 84 months of devel-
opment. We can also calculate the estimators I?~,,I,and s:l, for the distri-
bution of ultimate values of those same claims. Using regression we can
find constants which best fit
muI1= (I + fWFlxj NIld (4.1)
7
.Suh = (’ + d&j (4.2)
for the “mature” years. These parameters, along with the estimators ,niJ
and SE: for the current distribution of claims for accident year 1981 as
of December 31, 1988, yield the following estimates of the parameters
for the ultimate distribution of currently reported and open claims for
accident year I98 I :
rnzl, = (1 + I%& and (4.3)
s:; = (’ + ‘1s~:. (4.4)

Exhibits 1 and 2 provide a numerical example of this approach using

purely hypothetical data. In these exhibits. we assume losses for the first
 VARIABILITY IN LOSS RESERVES 89

seven accident years are sufficiently developed so that we “know” their

ultimate distributions and wish to estimate the distribution for accident
year 1981.

The distributions of claims reported at 84 months are shown in

Exhibit 1. Also shown in Exhibit I are the ultimate distributions for the
claims reported at 84 months for the first seven accident years, as well
as the corresponding parameters from the fitted lognormal distributions.
Exhibit 2 shows the results of the regression and corresponding
constants (u and c above) and coefficients (b and d above). Given the
lack of significance of the coefficient in the fit for s2, we assume no
relationship between s;: and s::. We thus use the sample mean and
variance for the ultimate distributions for the first seven years as our
parameter estimates. The bottom portion of Exhibit 2 then shows the
resulting estimates for the parameters of the ultimate distribution for
accident year 198 1.

At this point, other analyses (e.g., usual reserve estimation tech-

niques) could be used to modify these parameters to reflect the results
of those projections. It is a property of the lognormal distribution that
the coefficient of variation (ratio of the standard deviation to the mean)
can be expressed only in terms of the parameter s2:

c.v.z
= exp(.?)- 1. (4.5)
Thus, adjustments made to the rn:rt parameter will affect the mean
of the final distribution but not its relative variation, as measured by the
coefficient of variation. This technique does, however, have the benefit
of incorporating information regarding the current distribution of open
and reported claims in deriving the estimate for the ultimate distribution
of those claims.

We note that there is no chance of zero claims in the lognormal

distribution. If we were to use only that distribution as a model for
reported claims, then, strictly speaking, the number of claims is not
certain, for there may be open claims that will close without payment.
This can also be overcome by estimating the portion of those claims
which will close with payment separately, possibly also with the use of
regression.
90 VARIABII.ITY IN I OSS RESERVFS

Distributions for IBNR Reserves-One Approuch

For estimating the distribution of IBNR reserves, both the claim
counts and severity are uncertain. The parameters for the claim size
distribution could be considered in light of the ultimate value of claims
for more “mature” years which were reported after 84 months. The trend
in those costs could also be considered in selecting the distribution of
claim sizes.
One approach to estimating the distribution of claim counts would
be to assume it is Poisson and estimate the expected number of IBNR
claims using usual actuarial projection methods. Another approach, sim-
ilar to that used by Weissner [27], considers the reporting emergence as
a statistical distribution with known data truncated from above. Maxi-
mum likelihood estimators are then used to estimate the parameters of
that distribution. A beneft of this approach is that it can result in
estimates of both the mean and variance of the claim count distribution.
This approach begins by postulating a development curve in the form
of a probability distribution and then uses maximum likelihood estimators
along with known reported claims to estimate the ultimate number of
reported claims as well as an approximate distribution of that ultimate.
Though the application is in terms of reported claims, there is no inherent
reason that the same approach cannot be used to estimate the distribution
of ultimate losses directly.
We first assume that the number of claims reported through time t
can be expressed as
iJF(t;@. (4.6)
Here U is the (unknown) ultimate number of claims, and F(t$) is a
cumulative distribution function with parameter(s) 0 representing the
percent of ultimate claims reported through time t.
In this application, we think of the number of claims reported in
time period i as a grouped sample containing ,f; points in the interval
(c. ,, c,) from the distribution. We can use methods described in [25]
to iteratively approximate the maximum likelihood estimator of the pa-
rameter(s) 8 given these k observations. To this end, detine
P,.(6) = [F(cI-:&F(c.,-- ,;@]lF(cx;@. (4.7)
 VARlABlLlTY IN LOSS RESERVES 91

Here cr- I and c,. are the endpoints of the interval containing the fr
observations. Let f* denote the total number of claims reported through
k time periods, that is,

Define A(6) to be the matrix composed of the elements

(4.9)

and let the vector S(6) have the elements

With these functions, which involve only first derivatives of the

cumulative probability function with respect to its parameters, iteratively
calculate
6, = iin-, + [A(e,,-,)]-‘s(e,,-,). (4.11)
Now let h = F(ck;&) be the estimated percentage of claims reported
by time ck. The actual number of claims reported by time ck can then
be thought of as having a binomial distribution with (unknown) mean
Uh and variance Uh( 1 - h). Assume at this point that the binomial can
be approximated by a normal distribution. Thus, approximately,
T - N(Uh, Uh(1 - h)). (4.12)
Hence U = T/h is approximately normal:
U - N(f*lh, f(l - h)lh*). (4.13)
This results in an approximate distribution of IBNR claims I, where
+u-f- N(f*lh -f*, f*(l - h)lh*). (4.14)
Given these distribution estimates, an estimate of the distribution of
IBNR reserves for accident year 198 1 as of December 3 1, 1988 can then
be obtained. If it is assumed that this distribution and the distribution of
reserves for reported claims are stochastically independent, then an es-
timate of the distribution of total reserves can be made by convoluting
these two distributions.
92 VARIABILI rY IN I OSS RESERVkS

The assumption of independence may not be too restrictive in this

case. As of December 31, 1988. reported and IBNR claims form two
distinct populations. It is unlikely that fluctuations in the loss amounts
for a fixed number of known claims will lead to fluctuations in the
amounts, or counts, of claims yet to be reported. This does not, however,
address the question of parameter estimation for these populations and
the potential interrelationships there.
As an example of this approach, Exhibit 3 shows a hypothetical
claim emergence pattern for the first 84 months of development. We
selected a Weibull distribution to model this claims emergence. That
distribution’s cumulative density function can be written as
F(~;0~,8~) = 1 - exp{-exp[81 In(x/Oz)]}. (4.15)
The methods from [25] were then used to derive the parameter
estimates shown in Exhibit 3. These parameters result in an h-value of
0.930, with the resulting estimate of the expected number of IBNR
claims of 137 with a variance of 147.46.

Cornbinution of Years

The above calculations lead to an estimate of the distribution of total

reserves for a single accident year, in this case 1981. Though not ex-
plicitly stated, in practice they would probably be calculated for a single
coverage or line of insurance. For a multiple line company, however,
the distribution of total reserves, for all lines and for all years, is of
concern.
If one assumes that the distributions for the various lines of business
and accident years are all stochastically independent, the distribution of
total reserves could be estimated by convoluting the distributions for
individual lines and accident years. In some situations, the assumption
of independence may not be too restrictive.
In other situations, however, the reserve distributions for various
lines may not be independent; for example, in the bodily injury and
property damage portions of automobile liability coverage, some corre-
lations may sometimes be expected, especially in the distributions of the
number of claims.
There has been some activity in extending the Collective Risk Model
to include such interrelated events. Cummins and Wiltbank in [28] and
 VARIABILITY IN LOSS RESERVES 93

[29] consider multivariate models for claim count and size distributions.
These models can be thought of as considering the distribution of claims
arising from potentially different, but not independent perils. The paper
in [28] specifically addresses the automobile liability situation noted
above.

5. OTHER AREAS OF UNCERTAINTY

The applications discussed thus far have only addressed one area of
uncertainty, the statistical “noise” inherent in the insurance process,
assuming that ull distributions are correct. Not yet addressed are other
areas of uncertainty regarding the loss reserve estimates, such as:

1. How close are the selected parameters to the “real” parameters?

2. Are the distributions used in the model correct?
3. Is the Collective Risk Model the right one to use?
None of these questions has been answered yet, nor has the uncer-
tainty they imply been incorporated in the estimated distribution of
reserves. The first question, that regarding parameter uncertainty, is
sufficiently significant as to be the topic of a paper by Meyers and
Schenker [30]. In some situations, the variation due to parameter uncer-
tainty can outweigh the variation from the pure Collective Risk Model
itself. Needless to say, this should be recognized in any application of
the Collective Risk Model.

Also recognizing the importance of parameter uncertainty, Patrik and

John in [ 131reserve the term “Collective Risk Model” to a generalization
of what we present here. That generalization recognizes parameter un-
certainty by considering the parameters themselves as randomly drawn
from some probability space.

Often, parameter uncertainty is recognized by “expanding” the var-

iability of the component claim count or size distributions. If data is
lacking, such judgmental approaches may be all that is possible.

The possible approaches included above ( “Distributions for Reported

Claim Sizes,” . . . , “Distributions for IBNR Reserves”} lend themselves
for inclusion of parameter uncertainty. In the claim size distribution
estimates for reported claims, the parameter m:l, is estimated using linear
94 VARIABILI’I’Y IN L.OSS RESERVES

regression. Usual regression theory leads to the conclusion that the

variance of rniII can be expressed as
$ = (n - 2)Sf+(n - 41, (5.1)
where n is the number of points used in estimating the tit. and SE, is the
standard error of the forecast given the observed value for mk (“Distri-
butions for Reported Cluim Sixs” ).
We now assume that the claim size distribution is lognormal with
parameters m* and s:t, where m* is now unknown, but having a normal
distribution with mean rn:,, and variance SF. In this case, the final claim
size distribution will again be lognormal with parameters m:l, and
s:?~ + sf. Thus, the uncertainty regarding the scale parameter rnk is
translated to a widening of the coefficient of variation of the original
distribution. Other such “mixings” of distributions can be found in [31].
The bottom portion of Exhibit 2 continues with the example presented
above ( “Distributions fiw Reportcdd Cluim Sizes” ). For example, for
accident year 198 1, the standard deviation of the forecast of mk is 0.136
while the fitted SEEis 1.921. This results in an adjusted parameter of
1.939 for use with the lognormal distribution.
The maximum likelihood estimator methods presented in [25], as
outlined above (“Distributions,fcw lBNR Reserves” ), also provide means
to estimate the distribution of those estimators. What follows uses the
notation of Section 4 (“Distributions for IBNR Reseri,es”) and is an
application of those methods based on an unpublished presentation made
by Gary Venter.
Under suitable restrictions on the cumulative distribution function F,
the values of Cl,,,given in (4.J 1) converge-to the maximum likelihood
estimators of the parameters 8, call them t&. Also under suitable con-
ditions, the resulting parameters have a joint_ly normal distribution with
mean &I and variance-covariancc matrix [A(&)] ‘.
Now, h = F(cx;&) is a function of the maximum likelihood estimators
6,) and, following [25, pages 117-l 181 has an approximate normal dis-
tribution with mean 11,)= F(c~;$‘,). where 6,: denotes the estimate of the
maximum likelihood estimator IL. Apprcximately, then.
Tjh - N(Uho, c/Ml - ht,)). (5.2)
 VARIABILITY IN LOSS RESERVES 95

The variance of h can be approximated as

Var(h) = ,Jg, a,(%) $ (6;) g (6;). (5.3)

I J

Her: U;j denotes the i, j element of the approximate covariance matrix

[A(W)]- I. Thus, approximately,
T - N(Uho, Uho( 1 - ho) + u’Var(h)). (5.4)
Taking now
Uo = f*/ho (5.5)
as an estimate of the expected value of ultimate claims U, then, approx-
imately,
U = T/ho - N(Clo, [Uohdl - ho) + iJ: Var(h)llhz). (5.6)
There are admittedly many approximations in this estimation process.
It does, however, attempt to directly recognize the variability inherent
in the estimate of ultimate claims.
Using these approximations, the distribution of IBNR claims is then
approximately
I=U-f- N(fL - f*, [Cloho(l - ho) + Ug Var(h)llh$). (5.7)
When compared with formula (4.14), this indicates that parameter un-
certainty adds a factor of
iJi Var( h)lhi (5.8)
to the variance of the original unadjusted distribution. In the example in
Exhibit 3, Var(h) = 0.000135 for the fitted values of 0, and OZ. Also
shown in Exhibit 3 are the approximate parameter covariance matrix and
the partial derivatives used in calculating Var(h). In this case, the addi-
tional variance from (5.8) is 599.01, resulting in an indicated variance
in the projection of IBNR claims of 746.47.
In the examples presented here, we have used a single specific method
to estimate the parameters of the claim size distribution and distribution
of IBNR claims. Both of these methods are stochastic in nature and thus
supply information, under certain assumptions, regarding the uncertainty
inherent in their particular projections.
96 VARIABILI I’Y IN I OSS RESERVES

Usual actuarial projection methodology as described, for example,

by Skurnick [32] or Berquist and Sherman ]33] does not begin with an
underlying statistical model. Thus. the distribution of the projections
does not have a readily apparent statistical form. This problem is com-
pounded in practice where the actuary considers the results of several
different projection methods, often yielding different results. and selects
a best estimate of what the ultimate losses for a given coverage in a
given accident year will be.
As mentioned above, an approach used in these situations is to use
the best estimates of ultimate claim counts and severities as estimates of
E(N) and E(X) and then to select the claim count and size distributions
to have these expected values. Other parameter(s) are then selected to
represent the estimated variance in these two distributions and are derived
either by considering appropriate distributions of claims or judgmentally.
Paramctcr uncertainty may then be addressed by widening the resulting
distributions.
Thcsc methodologies do have the strength of addressing different
influences which may be apparent in the data. They also allow for the
introduction of seasoned judgment in interpreting the results of the pro-
jections or influences in the underlying data.
There arc also a variety of models which are statistically based.
Taylor’s work 1.341summarizes many different reserve estimation meth-
ods, and Ashc [22] provides a discussion of some of the work which
has been done to estimate variance in reserve projections using these
methods. Of particular note are regression-based methods of Taylor [ 35)
and Kalman Filter-based methods of DeJong and Zehnwirth 1361. Both
techniques look only to the historical development of losses for their
projections. It could be argued with this data that, put simply, “not all
that can happen has happened.” If that is true. thcsc methods may end
up understating the amount of variation in reserve projections. However,
they could be useful to quantify parameter uncertainty in the estimates
for the Collective Risk Model as presented here.
The answer to the question of how much uncertainty is added because
of the other two questions cited above is not nearly as clear as that for
parameter uncertainty. Estimates of parameter variability may address
some of the uncertainty inherent in the choice of a particular distribution
for the model. This may be further mitigated by reviewing the hts of
 VARIABILITY IN LOSS RESERVES 97

various distributions to the data available to minimize the chance of

picking the “wrong” one from a particular collection. However, it is
unlikely that, in actual applications, the second or third questions posed
above can be completely answered.
6. CONCLUSION

As can be seen from some of the questions raised, there appears to

be more work necessary to completely answer the question “How good
are our reserve estimates?” It has been the intent of this paper to present
an introduction to Collective Risk Theory for the first time reader, along
with a survey of some of the work that has been done which can be used
to attempt an answer to this question.
Without proper understanding, many tools can be misused. This is
true with Collective Risk Theory. The basic framework only addresses
certain portions of the potential variability in reserve estimates. Parameter
uncertainty is one significant area not specifically addressed by the basic
model; thus, it should be considered in any serious application to quan-
tifying reserve variability. Though some of the techniques outlined here
to address parameter uncertainty are necessarily complex and somewhat
abbreviated due to the intended scope of this paper, it is hoped that the
reader will appreciate the importance of this aspect of the Collective
Risk Model.
98 VARtABlLlTY IN LOSS RESERVES

REFERENCES

[l] “Statement of Principles Regarding Property and Casualty Loss and

Loss Adjustment Expense Reserves,” Casualty Actuarial Society,
May, 1988.
[2] S.P. D’Arcy, “Revisions in Loss Reserving Techniques Necessary
to Discount Property-Liability Loss Reserves,” PCAS LXXIV,
1987.
[3] C.D. Daykin, G.D. Bernstein, S.M. Coutts, E.R.F. Devitt, G.B.
Hey, D.I.W. Reynolds, and P.D. Smith, “Assessing the Solvency
and Financial Strength of a General Insurance Company,” Journal
of the Institute of Actuaries, I 13, 1987. p. 1.
[4] C.D. Daykin, G.D. Bernstein, S.M. Coutts, E.R.F. Devitt, G.B.
Hey, D.I.W. Reynolds, and P.D. Smith, “The Solvency of a Gen-
eral Insurance Company in Terms of Emerging Costs.” ASTZN
Bulletin, 17, 1987, p. I.
[5] R.E. Beard, T. Pentiktiinen, E. Pesonen, Risk Theory, The Sto-
chastic Basis uflnsurance, Third Edition. Chapman and Hall, 1984.
[6] K. Borch, Mathematical Theor?, @‘Insurance, D.C. Heath & Co.,
1970.
[7] H. Biihlmann, Mathematicd Methods of Risk Theory, Springer-
Verlag, 1970.
[8] H.L. Seal, Stochastic Theory of’ (1 Risk Bl4siness, Wiley, 1969.
[9] K. Borch, “Reformulation of Some Problems in the Theory of
Risk,” PCAS XLIX, 1962. p. 109.
[ lo] T. Pentiktiinen, “The Theory of Risk and Some Applications,”
Journal @‘Risk und /nsurtmce XLVII. 1980, p. 16.
[l l] A.L. Mayerson, D.A. Jones, and N. Bowers, Jr.. “The Credibility
of Pure Premiums,” PCAS LV, 1968, p. 175.
[ 121 P. Bratley, B.L. Fox, and L. E. Schrage, ,4 Cr4ide to Simulation,
Springer-Verlag , 1983.
[13] G.S. Patrik, and R.T. John, “Pricing Excess of Loss Casualty
Working Cover Reinsurance Treaties,” Casualty Actuarial Society
1980 Discussion Paper Program, p. 399.
 VARIABILITY IN LOSS RESERVES 99

[14] G.G. Venter, “Transformed Beta and Gamma Distributions and

Aggregate Losses,” PCAS LXX, 1983, p. 156.
[ 151 T. Pentikainen, “Approximative Evaluation of the Distribution
Function of Aggregate Claims,” ASTIN Bulletin 17, 1987, p. 15.
[ 161 Hon-Shiang Lau, “An Approach for Estimating the Aggregate Loss
of an Insurance Portfolio,” Journal of Risk and Insurance LI, 1984,
p. 20.
[ 171 S. Philbrick, “An Alternative Approach to Estimating Aggregate
Loss Distributions,” The Actuarial Review, v. 14, 1987, no. 3, p.
8.
[18] P.E. Heckman and G.G. Meyers, “The Calculation of Aggregate
Loss Distributions from Claim Severity and Claim Count Distri-
butions,” PCAS LXX, 1983, p. 22.
[ 191 H. Panjer, “Recursive Evaluation of a Family of Compound Dis-
tributions,” ASTIN Bulletin, 12, 1981, p. 22.
[20] B. Sundt, and W. Jewel], “Further Results on Recursive Evaluation
of Compound Distributions,” ASTIN Bulletin, 12, 1981, p. 27.
[21] G. Willmot, “Mixed Compound Poisson Distributions,” ASTIN
Bulletin, 16, 1986, p. S59.
[22] F. Ashe, “An Essay at Measuring the Variance of Estimates of
Outstanding Claim Payments,” ASTIN Bulletin, 16, 1986, p. S99.
[23] T. Pentikainen and J. Rantala, “Runoff Risk as a Part of Claims
Fluctuation,” ASTIN Bulletin, 16, 1986, p. 113.
[24] J.N. Stanard, “A Simulation Test of Prediction Errors of Loss
Reserve Estimation Techniques,” PCAS LXXII, 1985, p. 124.
[25] R.V. Hogg and S.A. Klugman, Loss Distributions, Wiley, 1984.
[26] J. Aitchison and J.A.C. Brown, The Lognormal Distribution With
Special Reference to its Uses in Economics, Cambridge University
Press, 1969.
[27] E.W. Weissner, “Estimation of the Distribution of Report Lags by
the Method of Maximum Likelihood,” PCAS LXV, 1978, p. 1.
100 VARIABILITY IN I.OSS RESERVES

[28] J.D. Cummins and L.J. Wiltbank, “Estimating the Total Claims
Distribution Using Multivariate Frequency and Severity Distribu-
tions,” JourrwI oj’Risk and Insurance L, 1983, p. 377.
1291 J.D. Cummins and L.J. Wiltbank. “A Multivariate Model of the
Total Claims Process,” ASTIN Bulletin 14, 1984, p. 45.
[30] G.G. Meyers and N. Schenker, “Parameter Uncertainty in the Col-
lective Risk Model,” PCAS LXX, 1983, p. Il.
[31] G.G. Venter, “Scale Parameter Mixing with Heavy Tailed Loss
Distributions,” Unpublished. 1985.
1321 D. Skumick, “A Survey of Loss Reserving Methods.” PCAS LX,
1973, p. 16.
[33] J.R. Berquist and R.E. Sherman, “Loss Reserve Adequacy Testing:
A Comprehensive, Systematic Approach.” PCAS LXIV, 1977, p.
123.
[ 34) G .C. Taylor, Claim Reserving in Non-l$) Irzsldr(lrzce, North-Hol-
land. 1986.
[35] G.C. Taylor, “Regression Models in Claims Analysis I: Theory,”
PCAS LXXIV, 1987, p. 354.
[36] P. DeJong and B. Zehnwirth, “Claims Reserving. State-space Mod-
els and the Kalman Filter,” Journul of the Institute qf Actuaries,
110, 1982, p. 157.
 VARIABILITY IN LOSSRESERVES 101

EXHIBIT 1
SHEET 1

DISTRIBUTION OF LOSSESFOR CLAIMS REPORTED

BY 84 MONTHS OF DEVELOPMENT

ACCIDENT YEAR I

At 84 Months Ultimate
Number Number
of Average of Average
Claim Size Range Claims CO9 Claims cost

$0 - $1,000 199 $450 170 $479

1,001 - 5,000 163 2,730 150 2,738
5,001 - 10,000 55 7,366 65 6,866
10.001 - 25,000 48 17,074 63 16,606
25,001 - 50,000 19 36,052 25 37,506
50,001 - 100,000 IO 71,898 I5 74,917
100.001 - 250.000 5 158.696 9 162,010
250,001 - 500,000 1 369,0 I 8 2 341,595
500,001 - 1.000.000 0 - I 71 1,158
I ,000,001 - 0 - 0 -
Total 500 500

Parameters of Fitted Lognormal Distributions

m 7.396 7.740
s-syuur-ed I .848 1.937
102 VARIABILITY IN LOSS RESERVES

EXHIBIT I
SHEET 2

DK~RIBIJTION OF LOSSESFORCLAIMS RWQRTED

BY 84 MONTHS OF DEVELOPMENT

ACCIDENT YEAR 2

At 84 Months Ultimate
Number Number
01 Average ol Average
Claim Size Range Claims cost Claims cost

$0 - $1,000 16X $443 150 $442

1,001 - 5,000 168 2,477 172 2,585
5,001 - 10,000 65 7,327 62 7,252
10,001 - 25 .OOO 59 15,55 I 64 17,055
25,001 - 50,000 25 37,613 29 35,638
50,001 - 100,000 14 72,826 I8 71,916
100.001 - 250,000 8 170,667 II 160,023
250,001 - 500,000 2 351.781 3 356,221
500,001 - 1,000,000 I 699,609 I 702,665
1,000,001 - 0 - 0
Total 510 5I0

Parameters of Fitted Lognormal Distributions

m 7.736 7.918
s-squared I.862 1.880
 VARIABILITY IN LOSSRESERVES 103

EXHIBIT 1
SHEET 3

DISTRIBUTION OF LOSSES FOR CLAIMS REPORTED

BY 84 MONTHS OF DEVELOPMENT

ACCIDENT YEAR 3

At 84 Months Ultimate
Number Number
of Average of Average
Claim Size Range Claims cost Claims cost

$0 - $1,000 172 $415 166 $445

1,001 - 5,000 167 2,502 173 2.622
5,001 - 10,000 62 7, I72 61 7,522
10,001 - 25,000 65 15,775 61 15,177
25,001 - 50,000 27 38,563 31 37,408
50,001 - 100,000 I5 74,796 I5 65,545
100,001 - 250,000 9 167,488 9 160,523
250,001 - 500,000 2 363,088 3 365,688
500,001 - I ,ooo,oOO I 663,006 1 705,967
I ,ooo,oo I - 0 0
Total 520 520

Parameters of Fitted Lognormal Distributions

7.754 7.797
1.899 I .896
104 VARIABILITY IN I.OS.5RISERVtS

EXHIBIT 1
SHt:kI 4

DISTRIBUTION OF LWSES FOR CLAIMS REPORTED

BY 84 MON.I-HS OF DEL’EI OPM~;N'T

ACCIDEN r Y~.AR 4

At X4 Months Ultimate
Number Number
of Average of Average
Claim Size Range Claims cost Claims cost

$0 - $1,000 160 $480 I61 $44 I

I.001 - 5,000 170 2,558 170 2,837
5,001 - 10,000 74 6.886 70 7,456
10,001 - 25,000 65 15,519 67 15,832
25,001 - 50,000 32 36,991 30 35,140
50,001 - 100,000 17 74.283 17 73,015
100.001 - 250,000 Y 163.701 II 158,295
250,001 - 500,000 3 370.993 3 345,297
500,001 - 1,000.000 I 720.3 I6 I 702,860
1,000,001 - 0 I 2, I 17,652
Total 531 531

Parameters of Fitted Lognormal Distributions

7.X96 7.897
I .X62 I.917
 VARIABILITY IN LOSSRESERVES 105

EXHIBIT 1
SHEET 5

DISTRIBUTION OF LOSSES FOR CLAIMS REPORTED

BY 84 MONTHS OF DEVELOPMENT

ACCIDENT YEAR 5

At 84 Months Ultimate
Number Number
of Average of Average
Claim Size Range Claims cost Claims cost

$0 - $ I ,000 I51 $443 140 $478

1,001 - 5,000 I77 2,647 172 2,531
5,001 - 10,000 73 7,809 74 7,858
10,001 - 25,000 71 16,229 73 16,888
25,001 - 50,000 34 35,331 39 32,982
50,001 - 100,000 20 72,039 21 72,711
100,001 - 250,000 II 153,797 I5 154,762
250,001 - 500,000 3 363,043 4 335,047
500,001 - I ,ooo,ooo I 703,801 2 679,978
1,000,001 - 0 I I ,924,372
Total 541 541

Parameters of Fitted Lognormal Distributions

m 8.003 8.145
s-squared 1.849 I .919
106 VARIABILITY IN LOSSRESERVES

EXHIBIT 1
SHEET 6

DISTRIBUTION OF LOSSESFOR CLAIMS REPORTED

BY 84 MONI‘HS OF DEVELOP~IENT

AKIDENI YEAR 6

At 84 Months Ultimate
Number Number
of Average Of Average
Claim Size Range Claims cost Claims cost

$0 - $ I ,000 I53 $436 I51 $450

I.001 - 5,000 181 3.50x I70 2,695
5,001 - 10,000 71 7,373 70 7,270
10,001 - 25,000 78 16,035 7x 16,929
25,001 - 50,000 35 37,119 37 36,601
50,001 - 100,000 19 77.169 23 68,545
100,001 - 250,000 II 157,721 15 164,521
250,001 - 500,000 3 366,860 5 337,331
500,001 - 1,000,000 I 716.312 2 694,022
I ,ooo.oo I - 0 I 2,312,174
Total 552 552

Parameters of Fitted Lognormal Distributions

m 8.012 8.103
s-squared I .x39 1.975
 VARIABILITY IN LOSSRESERVES 107

EXHIBIT 1
SHEET 7

DISTRIBUTION OF LOSSES FOR CLAIMS REPORTED

BY 84 MONTHS OF DEVELOPMENT

ACCIDENT YEAR 7

At 84 Months Ultimate
Number Number
of Average of Average
Claim Size Range Claims cost Claims cost

$0 - $1,000 140 $466 149 $468

1,001 - 5,000 187 2,697 I83 2,587
5,001 - 10,000 72 7,144 72 8,010
10,001 - 25,000 74 15,859 77 16,430
25,001 - 50,000 42 38,555 37 34,613
50,001 - 100,000 26 73,586 23 72,933
100,001 - 250,000 I4 158,619 I5 156,530
250,001 - 500,000 5 364,353 4 343,411
500,001 - I ,ooo,ooo 2 721,218 2 736,468
1,000,001 - I 2,128,700 - 1 2,045,068
Total 563 563

Parameters of Fitted Lognormal Distributions

m 8.180 8.105
s-squared I .909 1.923
10X VARIARII.ITY IN I OSS RESkRVI5

EXHIBIT I
SHEl:I’ 8

DISTKIBUTION OF LOSSESFOR CI AIMS REPORI’EI>

BY 84 MON OHS01: DEVF.IOPMI:W

Accident Year I Y8 I
Number
of Average
Claim Size Range Claims Cost

$0 - $ I,000 IIX $45’)

1,001 - 5 ,ooo IX3 2,707
5,001 - 10,000 x4 7,Y4Y
10,001 - 25.000 86 17.1 I4
2S,ool - 50.000 51 3.5.679
50,001 - 100,000 26 72,272
100,001 - 250.000 I7 151.062
250,001 - 500.000 6 366.2YY
500,001 - I .ooo.ooo 2 685,736
1,000,001 - 1 2.126.918
Total 456

Parameters of Fitted Lognormal Dictributions

111 8.41 I
s-srpured I.835
VARIABILITY IN LOSS RESERVES 109
110 VARIABILITY IN LOSS KESERVES

EXHIBIT 3

EXAMPLE CALCULATION OF CI.AIM COUNT

EXPECTED VALUE AND APPROXIMATE VARIANCE

Months of Reported Fitted Parameters (Weibull)

Development Claims
Theta( 1) = 1.195
Oto 12 463 Theta(2) = 37.077
12 to 24 382
24 to 36 369 Approximate Parameter
36 to 48 236 Covariance Matrix
48 to 60 198 (Inverse of A(Theta))
60 to 72 100
0.00145 -0.02309
72 to 84 74
-0.02309 1.63535
Total Reported 1,822
Partial Derivatives
of h with respect to
Theta( 1) 0.152
Theta(2) -0.00601
Var(h) - 0.000135
Additional Variance from
Parameter Uncertainty
599.01
h = 0.930
E(U) = 1,959
Var(U) - 147.46 (Unadjusted for parameter uncertainty)
Var(U) - 746.47 (Adjusted for parameter uncertainty)